File size: 16,847 Bytes
8f24287
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
"""
ML Training Optimizer Environment Implementation.

An MCP environment where an AI agent optimizes ML model training by:
- Configuring hyperparameters (optimizer, LR, batch size, etc.)
- Running real PyTorch training epochs
- Observing training curves and metrics
- Adjusting strategy based on convergence signals
- Submitting the best model for grading

All training is REAL — actual forward/backward passes with PyTorch on CPU.
"""

from typing import Any, Optional
from uuid import uuid4

from fastmcp import FastMCP

try:
    from openenv.core.env_server.mcp_environment import MCPEnvironment
    from openenv.core.env_server.types import Action, Observation, State
except ImportError:
    from openenv.core.env_server.mcp_environment import MCPEnvironment
    from openenv.core.env_server.types import Action, Observation, State

from .trainer import Trainer, TrainingConfig
from .tasks import TASKS, grade_task


class MLTrainerEnvironment(MCPEnvironment):
    """
    An OpenEnv environment for ML training optimization.

    The agent interacts via MCP tools to configure and run real PyTorch
    model training, observing actual training dynamics and making decisions
    to optimize performance.
    """

    SUPPORTS_CONCURRENT_SESSIONS: bool = True

    def __init__(self):
        """Initialize the ML trainer environment with MCP tools."""
        mcp = FastMCP("ml_trainer_env")
        self._setup_tools(mcp)
        super().__init__(mcp)

        self._state = State(episode_id=str(uuid4()), step_count=0)
        self._task_id: Optional[str] = None
        self._trainer: Optional[Trainer] = None
        self._max_steps = 15
        self._submitted = False
        self._prev_best_val_acc = 0.0
        self._cumulative_reward = 0.0

    def _setup_tools(self, mcp: FastMCP):
        """Register all MCP tools for the environment."""

        @mcp.tool
        def configure_training(
            optimizer: str = "adam",
            learning_rate: float = 0.001,
            batch_size: int = 64,
            weight_decay: float = 0.0,
            dropout: float = 0.0,
            lr_schedule: str = "constant",
            warmup_epochs: int = 5,
            augmentation: bool = False,
            augmentation_strength: float = 0.5,
        ) -> dict:
            """
            Configure the training hyperparameters. Call this before running
            training or to change the configuration mid-training.

            Note: Changing dropout will reset the model and training progress.
            Other changes (optimizer, LR, batch size, etc.) are applied without
            resetting the model.

            Args:
                optimizer: Optimizer to use. One of: 'sgd', 'adam', 'adamw'
                learning_rate: Learning rate (e.g., 0.001, 0.01, 0.1)
                batch_size: Batch size for training (e.g., 32, 64, 128)
                weight_decay: L2 regularization strength (e.g., 0.0, 0.0001, 0.001)
                dropout: Dropout rate for the model (e.g., 0.0, 0.2, 0.5).
                    WARNING: changing dropout resets the model!
                lr_schedule: Learning rate schedule. One of: 'constant', 'step',
                    'cosine', 'warmup_cosine'
                warmup_epochs: Number of warmup epochs (only for 'warmup_cosine')
                augmentation: Whether to enable data augmentation
                augmentation_strength: Augmentation intensity (0.0 to 1.0)

            Returns:
                Status message and current training metrics
            """
            if self._trainer is None:
                return {"error": "No task loaded. Call reset() first."}
            if self._submitted:
                return {"error": "Model already submitted. Call reset() to start a new episode."}

            config = TrainingConfig(
                optimizer=optimizer,
                learning_rate=learning_rate,
                batch_size=batch_size,
                weight_decay=weight_decay,
                dropout=dropout,
                lr_schedule=lr_schedule,
                warmup_epochs=warmup_epochs,
                augmentation=augmentation,
                augmentation_strength=augmentation_strength,
            )
            status = self._trainer.configure(config)
            metrics = self._trainer.get_metrics_summary()
            return {"status": status, "metrics": metrics}

        @mcp.tool
        def run_epochs(num_epochs: int = 10) -> dict:
            """
            Run N epochs of real PyTorch training with the current configuration.
            Each epoch performs actual forward/backward passes on the training data
            and evaluates on the validation set.

            Training time per epoch varies by task:
            - Easy (MNIST MLP): ~0.5-1 second
            - Medium (FashionMNIST CNN): ~1-2 seconds
            - Hard (CIFAR-10 CNN): ~2-3 seconds

            Args:
                num_epochs: Number of epochs to train (1-20, will be clamped
                    to remaining budget)

            Returns:
                Training metrics after running, including loss, accuracy,
                convergence signal, and remaining budget
            """
            if self._trainer is None:
                return {"error": "No task loaded. Call reset() first."}
            if self._submitted:
                return {"error": "Model already submitted."}
            if not self._trainer._initialized:
                return {"error": "Training not configured. Call configure_training first."}

            num_epochs = max(1, min(20, num_epochs))
            epoch_metrics = self._trainer.train_epochs(num_epochs)

            metrics = self._trainer.get_metrics_summary()
            metrics["epochs_just_trained"] = len(epoch_metrics)
            if epoch_metrics:
                metrics["epoch_time_seconds"] = round(epoch_metrics[-1].epoch_time_seconds, 2)

            return metrics

        @mcp.tool
        def adjust_learning_rate(new_lr: float) -> dict:
            """
            Adjust the learning rate without changing other settings.
            Useful for manual LR scheduling or responding to plateaus.

            Args:
                new_lr: New learning rate value (e.g., 0.0001)

            Returns:
                Status message and current metrics
            """
            if self._trainer is None:
                return {"error": "No task loaded. Call reset() first."}
            if self._submitted:
                return {"error": "Model already submitted."}

            status = self._trainer.adjust_learning_rate(new_lr)
            metrics = self._trainer.get_metrics_summary()
            return {"status": status, "metrics": metrics}

        @mcp.tool
        def toggle_augmentation(enabled: bool = True, strength: float = 0.5) -> dict:
            """
            Toggle data augmentation on/off and set its strength.
            Augmentation adds random transformations to training data to
            reduce overfitting (random rotations, flips, crops, color jitter).

            Args:
                enabled: Whether to enable augmentation
                strength: Augmentation intensity from 0.0 (minimal) to 1.0 (strong)

            Returns:
                Status message and current metrics
            """
            if self._trainer is None:
                return {"error": "No task loaded. Call reset() first."}
            if self._submitted:
                return {"error": "Model already submitted."}
            if self._trainer.state.config is None:
                return {"error": "Call configure_training first."}

            config = self._trainer.state.config
            config.augmentation = enabled
            config.augmentation_strength = max(0.0, min(1.0, strength))
            self._trainer.configure(config)
            metrics = self._trainer.get_metrics_summary()
            return {
                "status": f"Augmentation {'enabled' if enabled else 'disabled'} (strength={strength})",
                "metrics": metrics,
            }

        @mcp.tool
        def get_training_status() -> dict:
            """
            Get the current training status without advancing training.
            Returns all metrics, convergence signal, and remaining budget.

            Returns:
                Current training metrics and configuration
            """
            if self._trainer is None:
                return {"error": "No task loaded. Call reset() first."}

            metrics = self._trainer.get_metrics_summary()
            config = self._trainer.state.config
            if config:
                metrics["current_config"] = {
                    "optimizer": config.optimizer,
                    "learning_rate": config.learning_rate,
                    "batch_size": config.batch_size,
                    "weight_decay": config.weight_decay,
                    "dropout": config.dropout,
                    "lr_schedule": config.lr_schedule,
                    "augmentation": config.augmentation,
                    "augmentation_strength": config.augmentation_strength,
                }
            return metrics

        @mcp.tool
        def submit_model() -> dict:
            """
            Submit the current best model for final grading.
            This ends the episode and returns the final score (0.0 to 1.0).
            The score is based on the best validation accuracy achieved
            during the entire training run, not just the current epoch.

            Returns:
                Final grade with score and detailed breakdown
            """
            if self._trainer is None:
                return {"error": "No task loaded. Call reset() first."}
            if self._submitted:
                return {"error": "Model already submitted."}

            self._submitted = True
            grade = grade_task(self._task_id, self._trainer)
            metrics = self._trainer.get_metrics_summary()

            return {
                "status": "Model submitted for grading",
                "grade": grade,
                "final_metrics": metrics,
            }

    def reset(
        self,
        seed: Optional[int] = None,
        episode_id: Optional[str] = None,
        task_id: str = "easy_mnist",
        **kwargs: Any,
    ) -> Observation:
        """
        Reset the environment for a new training task.

        Args:
            seed: Optional random seed override
            episode_id: Optional episode ID
            task_id: Task to load: 'easy_mnist', 'medium_fashion', or 'hard_cifar'

        Returns:
            Observation with task description and initial state
        """
        if task_id not in TASKS:
            task_id = "easy_mnist"

        task = TASKS[task_id]
        self._task_id = task_id
        self._submitted = False
        self._prev_best_val_acc = 0.0
        self._cumulative_reward = 0.0

        self._state = State(
            episode_id=episode_id or str(uuid4()),
            step_count=0,
        )

        task_seed = seed if seed is not None else task.seed
        self._trainer = Trainer(
            model_type=task.model_type,
            dataset_name=task.dataset_name,
            max_epochs=task.max_epochs,
            seed=task_seed,
        )

        return Observation(
            done=False,
            reward=0.0,
            metadata={
                "status": "ready",
                "task_id": task_id,
                "task_name": task.name,
                "task_description": task.description,
                "difficulty": task.difficulty,
                "model_type": task.model_type,
                "dataset": task.dataset_name,
                "max_epochs": task.max_epochs,
                "max_steps": self._max_steps,
                "target_metric": task.target_metric,
                "target_value": task.target_value,
                "available_tools": [
                    "configure_training", "run_epochs", "adjust_learning_rate",
                    "toggle_augmentation", "get_training_status", "submit_model",
                ],
                "hint": (
                    "Start by calling configure_training to set hyperparameters, "
                    "then call run_epochs to train. Monitor metrics and adjust as needed. "
                    "Call submit_model when satisfied with performance."
                ),
            },
        )

    def _compute_step_reward(self) -> float:
        """Compute reward signal for the current step."""
        if self._trainer is None or not self._trainer.state.val_acc_history:
            return 0.0

        reward = 0.0
        current_best = self._trainer.state.best_val_accuracy
        current_val_acc = self._trainer.state.val_acc_history[-1]

        # 1. Progress reward: val_accuracy improvement
        if current_best > self._prev_best_val_acc:
            improvement = current_best - self._prev_best_val_acc
            reward += 0.3 * improvement
            self._prev_best_val_acc = current_best

        # 2. Loss decrease reward
        if len(self._trainer.state.val_loss_history) >= 2:
            prev_loss = self._trainer.state.val_loss_history[-2]
            curr_loss = self._trainer.state.val_loss_history[-1]
            if curr_loss < prev_loss:
                reward += 0.05

        # 3. Divergence penalty
        if self._trainer.state.is_diverged:
            reward -= 0.2

        # 4. Overfitting penalty
        gap = self._trainer.get_overfitting_gap()
        if gap > 0.08:
            reward -= 0.05 * (gap - 0.08)

        # 5. Wasted compute penalty (for hard task)
        if self._task_id == "hard_cifar":
            wasted = self._trainer.get_wasted_epochs()
            if wasted > 10:
                reward -= 0.02

        return round(reward, 4)

    def _step_impl(
        self,
        action: Action,
        timeout_s: Optional[float] = None,
        **kwargs: Any,
    ) -> Observation:
        """Handle non-MCP actions (return error)."""
        return Observation(
            done=False,
            reward=0.0,
            metadata={
                "error": f"Unknown action type: {type(action).__name__}. "
                "Use ListToolsAction or CallToolAction for MCP interactions."
            },
        )

    def step(
        self,
        action: Action,
        timeout_s: Optional[float] = None,
        **kwargs: Any,
    ) -> Observation:
        """Execute a step. Routes MCP actions through base class."""
        self._state.step_count += 1

        obs = super().step(action, timeout_s=timeout_s, **kwargs)

        # Compute reward and check termination
        reward = self._compute_step_reward()
        done = self._submitted or self._state.step_count >= self._max_steps

        # If max steps reached without submission, auto-submit
        if self._state.step_count >= self._max_steps and not self._submitted:
            self._submitted = True
            if self._trainer and self._trainer._initialized:
                grade = grade_task(self._task_id, self._trainer)
                reward += grade["score"]
                obs.metadata = obs.metadata or {}
                obs.metadata["auto_submitted"] = True
                obs.metadata["grade"] = grade

        # Add submission bonus to reward
        if self._submitted and self._trainer and self._trainer._initialized:
            # Check if this is the submission step
            obs_metadata = obs.metadata or {}
            if "grade" in obs_metadata:
                reward += obs_metadata["grade"].get("score", 0.0)

        obs.reward = reward
        obs.done = done
        self._cumulative_reward += reward

        return obs

    async def step_async(
        self,
        action: Action,
        timeout_s: Optional[float] = None,
        **kwargs: Any,
    ) -> Observation:
        """Async step for WebSocket handler."""
        self._state.step_count += 1
        obs = await super().step_async(action, timeout_s=timeout_s, **kwargs)

        reward = self._compute_step_reward()
        done = self._submitted or self._state.step_count >= self._max_steps

        if self._state.step_count >= self._max_steps and not self._submitted:
            self._submitted = True
            if self._trainer and self._trainer._initialized:
                grade = grade_task(self._task_id, self._trainer)
                reward += grade["score"]
                obs.metadata = obs.metadata or {}
                obs.metadata["auto_submitted"] = True
                obs.metadata["grade"] = grade

        obs.reward = reward
        obs.done = done
        self._cumulative_reward += reward

        return obs

    @property
    def state(self) -> State:
        """Get the current environment state."""
        return self._state